Modeling and maximizing information diffusion over hypergraphs based on deep reinforcement learning

Information diffusion is a hot research topic in the social computing field, which has a wide range of applications in economy, politics, and life. Most of the existing studies mainly focused on the information diffusion process in the ordinary graph, where the nodes representing users are connected by second-order relationships. However, the high-order relationships (e.g., coauthorship, membership) containing more than two nodes also exist in the real society, which can be captured well by the hypergraph. As the transformation from a high-order relation to several second-order relations will cause information loss, how to model the information diffusion process over hypergraph and how to maximize diffusion scale are two critical but still remained open problems. In this paper, we first propose the Crowd-LT model for hypergraph, which integrates bandwagon effect from psychology. Then the framework of Influence Maximization in Hypergraphs based on Deep Reinforcement Learning (IMH-DRL) is developed. It trains an agent to practice trial-and-error seed node selection tasks to gain an optimal strategy, which can be applied to diverse real-world hypergraphs. The extensive experiments on five hypergraphs prove the effectiveness of our proposed model and framework in terms of information spread and running time.